Limitations of MMSE in Cognitive Assessment: Revealing Latent Risk via Structural Brain Atrophy

Area of Science:

• Neuroscience and Artificial Intelligence
• Cognitive Impairment Biomarkers
• Explainable Deep Learning

Background:

• The Mini-Mental State Examination (MMSE) is a standard cognitive screening tool but has limitations in detecting early cognitive decline.
• Subtle cognitive changes may not be adequately captured by the MMSE, potentially missing early signs of neurodegeneration.
• Structural brain changes, like atrophy, are crucial indicators of cognitive health that complement cognitive test scores.

Purpose of the Study:

• To evaluate the combined utility of the MMSE and normalized Whole Brain Volume (nWBV) for classifying cognitive stages.
• To assess the independent and synergistic contributions of MMSE and nWBV in cognitive impairment risk screening.
• To investigate latent cognitive risk within the MMSE-normal population using structural brain atrophy indicators.

Main Methods:

• Developed an explainable deep-learning framework integrating MMSE scores and nWBV (a structural brain atrophy indicator).
• Compared four deep learning models (MLP, Tab ResNet, Tab Transformer, FT Transformer) using fivefold cross-validation.
• Conducted feature ablation analysis and employed Integrated Gradients (IG) and SHAP for interpretability to assess variable contributions.

Main Results:

• Model performance was driven by input data informativeness, not model complexity, with no significant differences between architectures.
• The MMSE showed strong discriminative power alone; however, nWBV significantly enhanced classification performance when combined with the MMSE.
• Interpretability analyses consistently identified MMSE and nWBV as key features, with stable results across cross-validation folds.

Conclusions:

• Integrating structural brain atrophy information (nWBV) with MMSE-centered cognitive assessments improves classification and risk screening.
• nWBV provides complementary structural risk signals, particularly valuable within the MMSE-normal subgroup for early detection.
• This explainable AI approach supports precision risk stratification and clinical decision-making in early cognitive decline.